@aws/ml-container-creator 0.4.0 → 0.5.0

package/templates/do/adapters/.gitkeep

@@ -0,0 +1,2 @@
+# This file ensures the do/adapters/ directory is tracked by git
+# Adapter metadata files (*.conf) are stored here after do/adapter add

package/templates/do/add-ic

@@ -0,0 +1,130 @@
+#!/bin/bash
+# Copyright Amazon.com, Inc. or its affiliates. All Rights Reserved.
+# SPDX-License-Identifier: Apache-2.0
+#
+# Add a new inference component to this project.
+# Creates a new IC config file in do/ic/ and deploys it immediately.
+
+set -e
+set -u
+set -o pipefail
+
+SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+source "${SCRIPT_DIR}/config"
+
+echo "➕ Add New Inference Component"
+echo "   Project: ${PROJECT_NAME}"
+echo ""
+
+# ============================================================
+# Prompt for IC name
+# ============================================================
+while true; do
+    read -p "IC name (lowercase alphanumeric + hyphens): " IC_NAME
+
+    # Validate: non-empty
+    if [ -z "${IC_NAME}" ]; then
+        echo "   ❌ IC name cannot be empty."
+        continue
+    fi
+
+    # Validate: lowercase alphanumeric + hyphens only
+    if ! echo "${IC_NAME}" | grep -qE '^[a-z0-9]([a-z0-9-]*[a-z0-9])?$'; then
+        echo "   ❌ IC name must be lowercase alphanumeric with hyphens (e.g., 'llama-70b')."
+        echo "      Must start and end with a letter or number."
+        continue
+    fi
+
+    # Validate: no collision with existing config
+    if [ -f "${SCRIPT_DIR}/ic/${IC_NAME}.conf" ]; then
+        echo "   ❌ IC config already exists: do/ic/${IC_NAME}.conf"
+        echo "      Choose a different name or edit the existing config."
+        continue
+    fi
+
+    break
+done
+
+# ============================================================
+# Prompt for image tag
+# ============================================================
+DEFAULT_IMAGE_TAG="${PROJECT_NAME}-latest"
+read -p "Image tag [${DEFAULT_IMAGE_TAG}]: " IC_IMAGE_TAG
+IC_IMAGE_TAG="${IC_IMAGE_TAG:-${DEFAULT_IMAGE_TAG}}"
+
+# ============================================================
+# Prompt for GPU count
+# ============================================================
+read -p "GPU count [1]: " IC_GPU_COUNT
+IC_GPU_COUNT="${IC_GPU_COUNT:-1}"
+
+# Validate numeric
+if ! echo "${IC_GPU_COUNT}" | grep -qE '^[0-9]+$'; then
+    echo "   ❌ GPU count must be a positive integer."
+    exit 1
+fi
+
+# ============================================================
+# Prompt for copy count
+# ============================================================
+read -p "Copy count [1]: " IC_COPY_COUNT
+IC_COPY_COUNT="${IC_COPY_COUNT:-1}"
+
+# Validate numeric
+if ! echo "${IC_COPY_COUNT}" | grep -qE '^[0-9]+$'; then
+    echo "   ❌ Copy count must be a positive integer."
+    exit 1
+fi
+
+# ============================================================
+# Prompt for memory MB
+# ============================================================
+read -p "Min memory MB [1024]: " IC_MIN_MEMORY_MB
+IC_MIN_MEMORY_MB="${IC_MIN_MEMORY_MB:-1024}"
+
+# Validate numeric
+if ! echo "${IC_MIN_MEMORY_MB}" | grep -qE '^[0-9]+$'; then
+    echo "   ❌ Memory MB must be a positive integer."
+    exit 1
+fi
+
+# ============================================================
+# Create IC config file
+# ============================================================
+IC_CONF_PATH="${SCRIPT_DIR}/ic/${IC_NAME}.conf"
+mkdir -p "${SCRIPT_DIR}/ic"
+
+cat > "${IC_CONF_PATH}" <<EOF
+# Per-IC configuration: ${IC_NAME}
+# Created by do/add-ic on $(date -u +"%Y-%m-%dT%H:%M:%SZ")
+#
+# This file is sourced by do/lib/inference-component.sh during deployment.
+# After deployment, IC_DEPLOYED_NAME and IC_DEPLOYED_AT will be appended
+# by the deploy script to track the active inference component.
+
+export IC_IMAGE_TAG="${IC_IMAGE_TAG}"
+export IC_GPU_COUNT=${IC_GPU_COUNT}
+export IC_COPY_COUNT=${IC_COPY_COUNT}
+export IC_MIN_MEMORY_MB=${IC_MIN_MEMORY_MB}
+export IC_STARTUP_TIMEOUT=900
+
+# Optional overrides:
+# export IC_MODEL_NAME="my-model-v2"
+# export IC_CONTAINER_ENV_EXTRA='"KEY":"value"'
+
+EOF
+
+echo ""
+echo "✅ Created IC config: do/ic/${IC_NAME}.conf"
+echo "   Image tag:  ${IC_IMAGE_TAG}"
+echo "   GPU count:  ${IC_GPU_COUNT}"
+echo "   Copy count: ${IC_COPY_COUNT}"
+echo "   Memory MB:  ${IC_MIN_MEMORY_MB}"
+echo ""
+
+# ============================================================
+# Deploy the new IC immediately
+# ============================================================
+echo "🚀 Deploying IC '${IC_NAME}'..."
+echo ""
+exec "${SCRIPT_DIR}/deploy" --ic "${IC_NAME}"

package/templates/do/benchmark

@@ -17,18 +17,29 @@ source "${SCRIPT_DIR}/config"
 # ── Parse flags ───────────────────────────────────────────────────────────────
 CLEAN_AFTER=false
 FORCE=false
-for arg in "$@"; do
-    case "$arg" in
-        --clean) CLEAN_AFTER=true ;;
-        --force) FORCE=true ;;
+IC_ARG=""
+ADAPTER_ARG=""
+while [ $# -gt 0 ]; do
+    case "$1" in
+        --clean) CLEAN_AFTER=true; shift ;;
+        --force) FORCE=true; shift ;;
+        --ic) shift; IC_ARG="${1:-}"; shift ;;
+        --adapter) shift; ADAPTER_ARG="${1:-}"; shift ;;
         --help|-h)
-            echo "Usage: ./do/benchmark [--force] [--clean]"
+            echo "Usage: ./do/benchmark [--ic <name>] [--adapter <name>] [--force] [--clean]"
             echo ""
             echo "Run SageMaker AI Benchmark against the deployed endpoint."
             echo ""
             echo "Options:"
-            echo "  --force    Create a new benchmark job even if one is already running"
-            echo "  --clean    Delete workload config and benchmark job after displaying results"
+            echo "  --ic <name>      Benchmark a specific inference component"
+            echo "  --adapter <name> Benchmark a specific LoRA adapter IC"
+            echo "  --force          Create a new benchmark job even if one is already running"
+            echo "  --clean          Delete workload config and benchmark job after displaying results"
+            echo ""
+            echo "IC resolution:"
+            echo "  --adapter <name> Use ADAPTER_IC_NAME from do/adapters/<name>.conf"
+            echo "  --ic <name>      Use IC_DEPLOYED_NAME from do/ic/<name>.conf"
+            echo "  (no flag)        Use first IC in do/ic/ alphabetically, or legacy config"
             echo ""
             echo "Idempotency:"
             echo "  If a benchmark job is already in progress, re-running without --force"
@@ -39,6 +50,7 @@ for arg in "$@"; do
             echo "  • AWS credentials must be configured"
             exit 0
             ;;
+        *) shift ;;
     esac
 done
 
@@ -52,6 +64,66 @@ if ! aws --version 2>&1 | grep -q "aws-cli/2"; then
     exit 1
 fi
 
+# ── Resolve inference component name ──────────────────────────────────────────
+# Resolution precedence: --adapter <name>, --ic <name>, first in do/ic/, or legacy config
+IC_NAME=""
+if [ -n "${ADAPTER_ARG}" ]; then
+    # Adapter name provided via --adapter flag — look up adapter IC
+    ADAPTER_CONF="${SCRIPT_DIR}/adapters/${ADAPTER_ARG}.conf"
+    if [ ! -f "${ADAPTER_CONF}" ]; then
+        echo "❌ Adapter config not found: do/adapters/${ADAPTER_ARG}.conf"
+        echo "   Available adapters:"
+        if [ -d "${SCRIPT_DIR}/adapters" ]; then
+            for conf in "${SCRIPT_DIR}"/adapters/*.conf; do
+                [ -f "${conf}" ] || continue
+                echo "     • $(basename "${conf}" .conf)"
+            done
+        else
+            echo "     (none)"
+        fi
+        exit 1
+    fi
+    ADAPTER_IC_NAME=""
+    source "${ADAPTER_CONF}"
+    if [ -z "${ADAPTER_IC_NAME}" ]; then
+        echo "❌ Adapter '${ADAPTER_ARG}' conf is missing ADAPTER_IC_NAME."
+        exit 1
+    fi
+    IC_NAME="${ADAPTER_IC_NAME}"
+elif [ -n "${IC_ARG}" ]; then
+    # Explicit IC name provided via --ic flag
+    IC_CONF="${SCRIPT_DIR}/ic/${IC_ARG}.conf"
+    if [ ! -f "${IC_CONF}" ]; then
+        echo "❌ IC config not found: do/ic/${IC_ARG}.conf"
+        exit 1
+    fi
+    IC_DEPLOYED_NAME=""
+    source "${IC_CONF}"
+    if [ -z "${IC_DEPLOYED_NAME}" ]; then
+        echo "❌ IC '${IC_ARG}' has not been deployed yet. Run ./do/deploy --ic ${IC_ARG} first."
+        exit 1
+    fi
+    IC_NAME="${IC_DEPLOYED_NAME}"
+elif [ -d "${SCRIPT_DIR}/ic" ]; then
+    # No --ic argument, but do/ic/ exists — use first IC alphabetically
+    for conf in "${SCRIPT_DIR}"/ic/*.conf; do
+        [ -f "${conf}" ] || continue
+        IC_DEPLOYED_NAME=""
+        source "${conf}"
+        if [ -n "${IC_DEPLOYED_NAME}" ]; then
+            IC_NAME="${IC_DEPLOYED_NAME}"
+            break
+        fi
+    done
+    if [ -z "${IC_NAME}" ]; then
+        echo "❌ No ICs deployed. Run ./do/deploy first."
+        exit 1
+    fi
+else
+    # Legacy: no do/ic/ directory, use INFERENCE_COMPONENT_NAME from do/config
+    IC_NAME="${INFERENCE_COMPONENT_NAME:-}"
+fi
+
 # ── Helper: update a variable in do/config ────────────────────────────────────
 _update_benchmark_var() {
     local var_name="$1"
@@ -123,7 +195,7 @@ MAX_POLL_ATTEMPTS=60  # 30 minutes max (60 * 30s)
 echo "📊 SageMaker AI Benchmark"
 echo "   Project: ${PROJECT_NAME}"
 echo "   Endpoint: ${ENDPOINT_NAME:-not set}"
-echo "   Inference Component: ${INFERENCE_COMPONENT_NAME:-not set}"
+echo "   Inference Component: ${IC_NAME:-not set}"
 echo "   Concurrency: ${BENCHMARK_CONCURRENCY}"
 echo "   Input tokens (mean): ${BENCHMARK_INPUT_TOKENS_MEAN}"
 echo "   Output tokens (mean): ${BENCHMARK_OUTPUT_TOKENS_MEAN}"
@@ -318,7 +390,7 @@ echo ""
 # Target the deployed endpoint and inference component with the workload config.
 echo "🚀 Step 2: Creating AI Benchmark Job: ${BENCHMARK_JOB_NAME}"
 
-BENCHMARK_TARGET="{\"Endpoint\":{\"Identifier\":\"${ENDPOINT_NAME}\",\"InferenceComponents\":[{\"Identifier\":\"${INFERENCE_COMPONENT_NAME}\"}]}}"
+BENCHMARK_TARGET="{\"Endpoint\":{\"Identifier\":\"${ENDPOINT_NAME}\",\"InferenceComponents\":[{\"Identifier\":\"${IC_NAME}\"}]}}"
 OUTPUT_CONFIG="{\"S3OutputLocation\":\"${BENCHMARK_S3_OUTPUT_PATH}\"}"
 
 if ! aws sagemaker create-ai-benchmark-job \